The Transparency-Revenue Conundrum in Social

Trading: Implications for Platforms and Investors

Mingwen Yang, Eric Zheng, and Vijay Mookerjee

Jindal School of Management, The University of Texas at Dallas

Abstract

Social trading – an emerging paradigm in the spirit of the sharing economy – enables

a trader to share her trading wisdom with other investors. A special type of social

trading is copy trading, where less experienced investors (followers) are allowed to

copy the trades of experts (traders) in real-time after paying a fee. Such a copy

trading mechanism often runs into a transparency-revenue conundrum. On the

one hand, social trading platforms need to release traders’ trades as transparently

as possible to allow followers to evaluate traders. On the other hand, complete

transparency may undercut the platform’s revenue since followers could free ride.

That is, followers could manually copy the delayed trades of a trader to circumvent

paying following fees.

This study addresses this simple, but fundamental problem by determining the

optimal policy for releasing trade information. Of key interest here is the real time

value of trade information. We measure this real time value using the concept of

profit-gap, i.e., the would-be profit difference between the real time and a delayed

execution of a trade. Our data shows that profit-gap increases with delay in a

concave manner. To provide empirical evidence of the conundrum, we explore how

the amount following is affected by delay and users’ evaluation activity on the

trader (measured by the number of views received by the trader’s profile page).

Surprisingly, when the number of views is high, a larger delay is associated with a

lower amount following. A large delay handicaps the ability of potential followers to

evaluate the trader, thus reducing the amount that would potentially have followed

1

the trader. This potential loss is higher when the number of views is higher. On

the other hand, a larger delay discourages existing followers to free ride and hence

protects the existing amount following a trader.

We propose the notion – money-at-risk – to quantify the possible loss of existing

followers (amount following) as a result of releasing delayed trades. The tradeoff

between transparency and money-at-risk is cast as an optimization problem that

attempts to maximize information transparency while respecting a money-at-risk

constraint. Three information release policies are compared: (1) Uniform Release

Policy, (2) Customized Money-at-Risk Release Policy, and (3) Customized Indiffer-

ence Policy. We demonstrate the performance of the above policies using data from

a leading social trading platform operating in the Foreign Exchange market. We

also study a Stochastic control formulation that directly optimizes platform revenue.

The control is the delay that is calculated as a function of the current amount of

money following a trader and the number of views received by the trader’s profile

page. Besides, the calculated revenue can be incorporated into the ranking algo-

rithm to provide a systematic way to infuse the platform’s goals into the ranking

the traders.

Keywords: Fintech; Social Trading ; Information Value; Information Release Pol-

icy

1 IntroductionThe emergence of new financial technologies (Fintech) has helped bridge the digital

divide of financial services, especially in terms of access to financial advising and wealth

management services. Among them is social trading, through which a retail investor

can manage her wealth (no matter how small it is) by directly following other traders’

financial advice (Eldridge 2017). These traders often share their opinions (aka financial

advice) on financial markets through specialized social media platforms (e. g., StockTwits

and Seeking Alpha), where traders can make friends, post their opinions, and directly

communicate with other investors (Doering et al. 2015). Investors can then use such

financial advice to make trading decisions.

Recently there emerges a more disruptive type of social trading platforms, represented

by eToro, Zulutrade and Collective2. They go beyond merely providing a platform for

traders to share their trading opinions, they even allow investors to observe and follow the

actual trading (action) of peer traders on stocks, currencies, and cryptocurrencies (Pelster

2

and Hofmann 2017). And we focus on these newer types of social trading services.

These social trading platforms have gained growing popularity as evidenced by re-

ceiving multiple rounds of venture capital funding (Reuters 2018) and the growing pool

of traders on these platforms with over 13.9 million online retail traders as of 2018 (Busi-

nessWire 2017), (BrokerNotes 2018). eToro alone has attracted over 9 million active

users in 2018, mostly small retail investors, who are allowed to open an account with as

little as 200 USD 1. The demand for social trading services is projected to explode in the

future (Empire 2017), covering 37% of the population of investors by 2021 (eToro 2017).

Most of these social trading platforms are regulated. For example, the Financial Con-

duct Authority (FCA) in 2015 stipulated a rule that all traders in social trading need to

comply with the MiFID II ruling to qualify as investment managers 2. A handful of these

platforms are allowed to operate in the U.S. market to serve U.S. residents, including

Collective2, Peertrade and Zulutrade’s Forex and FXCM markets 3.

Of central interest to us is the social trading feature called copy trading. This feature

enables investors to replicate – in real-time and in their own accounts – the actual trades of

other investors. An investor is allowed to follow other traders for a fee and have the plat-

form automatically execute the trades of other traders on her own account in a real time

manner. There are several key advantages of such a copy trading mechanism, compared

to other wealth management services: 1) More transparency since investors will observe

every single trade of a trader 4; 2 ) higher-level of control since the execution is done in

the investor’s own account5; 3) more reasonable fee (compared to funds), typically zero

management fee and 20 percent when there is a gain (https://www.zulutrade.com/trader-1https://bitcoinist.com/9-million-traders-flocked-etoros-social-trading-platform/2The key regulations on social trading include 1) No hedging of trades - a trader cannot have a trade

in the same instrument open in two opposite directions (Buy and Sell simultaneously); and 2) First In

First Out (FIFO) trades rule - when a trader has multiple trades open of the same pair in the same

direction, they must however be closed in the order they are opened, https://financefeeds.com/mifid-ii-

entering-age-completely-self-directed-traders-final-nail-goes-copy-trading-coffin/.3https://socialtradingguru.com/social-trading-for-us-residents4For example, although institutional money managers with $100 million or more in qualifying assets

are required to file quarterly an SEC Form 13F detailing their investment holdings, investors do not

know exactly when the fund manager bought and sold a commodity in the portfolio.5In institutional markets, investors have less flexibility over their portfolios; for example, less control

over timing the recognition of gains and no way to customize the specific set of commodities in the

portfolio.

3

guide)6, and 4) preventing potential manipulation that may occur in the financial advis-

ing industry, where financial advisors may strategically distort their recommendations by

“speaking in two tongues”, for example, issuing overly positive recommendations but less

optimistic forecasts (Malmendier and Shanthikumar 2014). Under copy trading, it is hard

for traders to “speak in two tongues” since their real investment actions are observed and

copied. Top social trading networks (platforms) that enable copy trading include eToro,

ZuluTrade, Ayondo, Tradeo, etc. For example, eToro launched this feature in 2007 for

followers to replicate traders’ trades (Kortekaas 2013). According to eToro, among its

300 billion U.S. dollars worth of trades, two-thirds were executed through copy trading

(Brand 2017).

We seek to devise a proper information revelation mechanism to optimize information

transparency in social trading. Improving information transparency has been one of

the key drivers underlying the current wave of financial technology innovations (Lee

and Shin 2018). Compared with traditional fund management, social trading platforms

provide much higher information transparency by sharing not only aggregated metrics

but detailed trade-level information among traders. Many social trading platforms have

gone to the extreme of publicizing all the trades of each trader (Glaser and Risius 2018).

This complete information transparency policy is one of the main drivers behind the

explosive growth of the user base in social trading (Röder and Walter 2017).

Information transparency in trading involves two aspects: (1) what information should

be disclosed (e.g., detailed versus aggregate trading performance), and (2) when informa-

tion should be disclosed (e.g., immediate versus delayed). Timing the release of informa-

tion is particularly relevant in social trading since trading information quickly loses value

in a fast-paced financial market. A trend in the social trading industry is to release as

much information to investors as possible (e.g. eToro and Zulutade have started releasing

each trade transacted by each trader including the securities bought or sold, time and6It is reported that investors earned an average of 4.67% on mutual funds over the last 20 years,

which is even 3.52% less than the average SP 500 index return; mutual funds provided an average return

of 6.92% over the last 5 years, around 3% less than the SP 500 index over the same period (Kim 2017).

As Warren Buffett described about fund managers: “Professionals in other fields, like dentists, bring a

lot to the layman, but people get nothing for their money from professional money managers.” The fund

industry is overdue for change. The fee war is just part of the problem; price-conscious investors also

want more transparency around how they pay for advice (Kapadia 2018).

4

price of the transaction etc. )7, and as such there is typically no ambiguity with regard

to decision 1 in social trading. Our study thus focuses on decision 2, when to release

information.

Specifically, concerning the decision of when to release information, in copy trading,

timing the release of information is crucial. Potential investors have to rely on the infor-

mation the platforms provide to evaluate traders in order to choose a subset of traders

to follow. Information becomes less transparent if it is released with longer delay. On

the other hand, releasing trade information with very little or no delay can also cause

problems. Copy trading is not free for followers: the platform functions as an online

broker and, in addition to a brokerage fee, it also charges followers a fee to follow the

trades of traders in real-time. Traders receive a commission calculated as a fraction of

the follower fees generated from their trades. Followers will not pay to follow if detailed

trade information is already released for free in real-time. Thus, traders will lose paid

followers and consequently, will lose the incentive to participate in the platform. There-

fore, a fundamental decision for a copy trading platform is to choose the right level of

information transparency with regard to the timing of releasing trade information.

At one extreme, some social trading platforms have allowed complete information

transparency by releasing real-time trading information of traders to the public.8 How-

ever, complete information transparency invokes the opportunistic behavior of investors,

since they no longer have to pay to follow traders and can get the real-time information

for free. At the other extreme, social trading platforms may choose to delay information

release indefinitely. This, of course, renders such information unusable. Most social trad-

ing platforms, however, seek a balance between complete information transparency and

complete hiding by adding some time delay to release detailed, trade-level information.

We therefore focus on the decision of determining the optimal time to release detailed7Concerning the decision of what information to release, most platforms release full information. By

providing the detailed trade-level information, the platform protects itself by delegating the responsibility

of evaluating traders to followers, thereby reducing the probability of followers making biased decisions

acting on aggregated performance indicators (e.g. monthly returns). However, because releasing detailed

information could increase the information processing burden of investors, detailed information is usually

made available on-demand (e.g., by clicking on a link), while only summarized information is made visible

by default.8ZuluTrade initially experimented with this policy around 2013, but later, moved to the policy of

releasing trade information with some delay.

5

trading information. While releasing such valuable trade information is important, dis-

playing this information with a large time delay may defeat its purpose. A large delay

makes it difficult for potential followers to evaluate traders thoroughly. Traders may trade

with different strategies on different securities and they often need to adjust their trading

strategies based upon the changing market conditions. Potential followers would only be

able to evaluate the on-going strategies if the trades are released without too much delay.

At the high level, we answer the simple but fundamental question: is copy trading

worthwhile for followers? In other words, how valuable is it for followers to know (in real-

time) the trades of traders, given that they must pay a fee to receive this information?

This is a fundamental question all copy trading platforms need to answer before designing

an optimal information revelation policy.

The answer to this question is vital for several reasons. First, the copy trading plat-

forms will collapse if followers do not see the real-time value in their service. Thus, the

answer will help followers decide when it is valuable to follow in real-time and when

it is not. Second, answering the above question is key for the platform to best mon-

etize the dissemination of its financial advice. Third, it helps platforms determine the

right magnitude of delay to add before releasing trade information to the public for free,

while preventing arbitrage opportunities and at the same time guaranteeing information

transparency. For example, ZuluTrade releases traders’ trades after 30 minutes to al-

low potential followers to evaluate different traders before following them. Lastly, the

answer offers cues for the platform to personalize the delay (and, possibly the following

fee) for each trader, based on the characteristics of the trader and the market conditions

associated with the investment product.

Thus, in this study, we specifically investigate the following three research questions.

• How does the level of information transparency (as measured by delay) affect the

profit of a trade?

• How is the amount following a trader affected by information transparency?

• How should the platform design an information revelation policy?

Empirically, we examine the copy trading phenomenon implemented by ZuluTrade,

a leading social trading platform that mainly operates in the Foreign Exchange market.

It mainly deals with day trading, and thus the time to open and close a trade within

6

a market is important. ZuluTrade provides two types of accounts: trader and follower.

A follower can “follow” one or more traders. Trades from trader accounts will be copied

in real-time and automatically executed in a follower’s own accounts. The platform

functions as an online broker and charges followers a fee (per trade with some rate) to

follow the trades of a trader. While traders themselves also pay a commission to the

platform to execute their trades, they receive a kickback bonus (usually settled on a

monthly basis) that is calculated based upon the amount of money that follows their

trades. The platform operates as a two-sided market, where more followers will motivate

more traders to join and vice versa. Followers pay to copy real-time trades; however, these

trades are available for free after some time of delay (for example, 30 minutes during our

study period). Thus, investors can free ride if they are willing to accept the time delay

introduced by the platform.

The data we obtained consists of individual traders’ trades executed on ZuluTrade,

for the Foreign Exchange (Forex, or currency trading) market. Our interest is to quantify

the information value of knowing these trades in real-time. We define profit-gap, as the

difference between the real-time profit of a trade and the simulated profit (calculated using

historical Forex spot price data) of the trade, but executed with some open delay and/or

some close delay. We use the data to examine how delay affects the profit-gap, after

controlling for various factors related to the trader and other market conditions. Next,

from the data, we present empirical evidence to demonstrate the platform’s dilemma:

Lowering the delay increases transparency (that could potentially increase future revenue)

but risks losing the amount of money following (and hence hurts the platform’s current

revenue).

We then formulate and solve several optimization problems that address the platform’s

transparency-revenue conundrum. First, we examine the current approach adopted by

Zulutrade and study how it can be improved. Currently ZuluTrade seeks to maximize

transparency while holding the Money-at-Risk at or below an acceptable level, where

Money-at-Risk measures the vulnerability of the platform’s current revenue as a function

of the delay and the fees the platform charges followers to receive information on a

trader’s trades in real-time. We study an improved information release policy that is

trader specific and demonstrate that it can substantially increase transparency (reduce

delay) while maintaining the Money-at-Risk at a level equal to ZuluTrade’s current release

7

policy.

We further study – using a Stochastic Optimal Control formulation of the problem –

how the information release policy can be chosen to optimize revenue. That is, rather

than indirectly addressing the problem by maximizing transparency subject to given

(exogenously chosen) Money-at-Risk, we set up a model to directly maximize platform

revenue. The Stochastic Control problem is solved to yield a feedback control (i.e.,

delay) that is based on the current amount following a trader and the current attention

being received by the trader from potential followers (measured by the trader’s profile

page views). The calculated revenue can be incorporated into the ranking algorithm to

provide a systematic way to infuse the platform’s goals into the ranking of the traders. In

so doing, this study also helps followers to determine whether it is worthwhile to follow

a specific trader in real-time.

The rest of this paper is organized as follows. We next provide a review of related

work. In Section 3, we introduce the data and the operations of the platform. In Section

4 we study the data to provide the empirical basis for the optimization models in Section

5 and Section 6. In Section 7, we provide a discussion and conclude the study.

2 Literature ReviewIn this section, we briefly summarize the literature on social trading. The first stream

of literature studies investment decisions made by individuals as a result of their social

interactions. Ammann and Schaub (2016) investigate the role of social interaction in

investment decisions by mining trader posts and other communications. They find that

traders with superior performance are more likely to discuss their investment strategies.

Heimer (2016) focuses on a phenomenon called the disposition effect. This anomaly,

discovered in behavioral finance, uncovers the tendency of investors selling too early in

the up market, while holding too long in a down market. Disposition effect is found to

be magnified when investors receive advice from their friends.

The second related stream of literature studies the phenomenon of copy trading. Do-

ering et al. (2015) describe how copy trading platforms are organized and discuss the

basic mechanics behind the relationship between signal provider (portfolio manager) and

signal followers (investors). They find that signal providers typically engage in active

trading rather than buy-and-hold strategies, which result in non-normal return distribu-

tion. Lee and Ma (2015) propose a system identifying traders with good and consistent

8

performance to answer the question “whom to follow”. Oehler et al. (2016) show that, on

average, traders on wikifolios (a copy trading platform) do not outperform the market on

average but the best performing traders earn significant short-term excess returns.

Within this literature on social trading, several studies analyzed the trading behavior

and performance of traders in copy trading. Pan et al. (2012) study the role of so-

cial mechanisms in a financial system and find that social trades outperform individual

trades. Röder and Walter (2017) discover that traders who communicate actively with

investors attract significantly more attention, and visibility of their trading portfolios

boosts investments. Breitmayer et al. (2017) investigate the trading patterns of traders

who received social recognition for their investment advice. They show that confirmatory

social recognition leads to increased trading activity. Pelster and Hofmann (2017) study

the relationship between providing financial advice and the disposition effect. They find

that leading traders are more susceptible to the disposition effect than traders without

followers.

The third stream of relevant literature is on information transparency with regard to

what information and how much information to release in financial services. For example,

in crowdfunding, platforms need to decide what level of borrower information to release

to help these lenders evaluate a loan. Crowdfunding studies find that what information

(explicit or implicit) to release strongly influence the overall market efficiency and the

lenders’ decisions to participate (Herzenstein et al. 2011, Mäschle 2012). Zhou et al.

(2018) build a structural model to uncover how lenders’ behaviors are affected by an

exogenous information-disclosure policy change. They show that displaying extra infor-

mation leads to a higher browsing propensity and helps lenders to make sound investment

decisions.

The fourth stream of related literature is on the value of information for devising

trading strategies. Mutual funds are mandated to disclose their portfolio holdings to

investors periodically, e.g. quarterly. Some investors might mimic the trading strategy

from the released portfolio, called copycat funds in finance. Verbeek and Wang (2013)

indicate that free-riding on disclosed fund holdings is an attractive strategy and suggest

that mutual funds may suffer from such information disclosure regulations. The time-

liness of portfolio holdings disclosure has been of interest among regulators, academics

and practitioners since the Investment Company Act of 1940. The Securities Exchange

9

Commission (SEC) has been trying to strike a balance (a uniform delay across all mutual

funds) between investors’ interest in timely disclosure and the potential costs associated

with revealing the strategies of investment managers (Hee Choi and Chhabria 2012). The

information required to be released in mutual funds is the portfolio holdings, not the ex-

act time when the mutual fund manager bought and sold the portfolio holdings. In this

study, our data comes from a social trading platform focusing on day trading and the

platform does releases the time traders execute a trade.

Different from the prior research that studies the social trading phenomenon mainly

from a social or behavioral perspective, our study is more normative. We address the

fundamental mechanism design problem in copy trading to determine the right level of

information to release to the public. Specifically, the main goal of this study is to measure

the economic value of paying for, and obtaining, real-time trade information on a copy

trading platform, and to address how to choose the optimal level of transparency for copy

trading platforms. None of the prior research has considered this perspective.

3 Data and Platform OperationsIn this section, we first introduce the data and then describe the operations of the

social trading platform we consider.

3.1 Data

Our data comes from ZuluTrade, one of the largest copy trading platforms in the

world. The platform allows followers to auto-copy Forex trades made by financial experts

(traders). Each trader owns a public profile page, which reveals information on her

past trading performance tracing back to the first day the trader joined the platform.

ZuluTrade releases various performance metrics including the total profit of all the trades

a trader has executed via the platform, the best and the worst realized profit of the trade

among all her trades, the percentage of winning trades, the number of followers following

the trader, the total amount of money following the trader, the number of weeks the

trader has been trading on the platform, the number of views the trader’s profile has

received, etc. Importantly, ZuluTrade also releases on-going trades (trades have been

opened but not closed yet) for free after adding some time delay to the public. Figure 1

provides a screen shot of a trader’s profile page.

We obtained individual trading information of 15,352 traders during a 17 month

period from August 2016 to December 2017. The data is at the most granular trade

10

Figure 1: Snapshot of a Trader’s Profile Page

level possible for each trader. The detailed trade information includes the currency (e.g.

EUR/USD), type (buy or sell), standard lot size, date open, date close, open price (the

spot price at the time when the trade was opened), close price (the spot price at the time

when the trade was closed), profit, the highest potential profit during the holding period

of the trade, the lowest potential profit (max drawdown) during the holding period of the

trade, etc. We focus on the intra-day trades since the Forex market is volatile and more

than half trades are open and closed within the same day. The Forex market provides

an ideal market (environment) to study the time value of trade information since the

profit of a trade is very sensitive to the magnitude of delay. In this study, we focus on

11

the five largest currency markets – EUR/USD, GBP/USD, GBP/JPY, USD/JPY, and

USD/CAD – that together account for more than 60% of all the trades on this platform.

Table 1 presents the proportions of the five currencies among all the intra-day trades.

Table 1: Relative Proportions of Five Currencies

Currency Frequency Percent

EUR/USD 585,939 37.83%

GBP/JPY 226,230 14.61%

GBP/USD 385,702 24.90%

USD/CAD 126,421 8.16%

USD/JPY 224,461 14.49%

Total 1,548,753 100%

The second dataset we obtain is the historical spot prices for these five currency

markets from the Dukascopy Historical Data Feed. We use this data set to calculate the

hypothetical profit (loss) of a trade, which we will elaborate in the next section.

3.2 ZuluTrade Operations

On the ZuluTrade platform, once a follower clicks to “follow” a trader, the trades

of this trader are automatically copied and executed in this follower’s account. The

platform charges followers commission (following fee): 2 pips for a complete open-close

trade per standard lot size 9 This commission is already factored in the buy and sell price.

Followers can choose how to copy traders’ trades. For example, followers can choose a

fraction γ ∈ [0, 1] for each trader; only a γ fraction of the trader’s trade amount will

be executed in the copied trade. The copying trades will automatically be executed in

real-time in the follower’s account. Followers who do not follow a trader cannot receive

this trader’s trade information in real-time. However, the platform releases this trade

information with some delay on the trader’s profile page for the public to see, and for

free.

The platform shares revenue with traders: 0.5 pip for each standard lot executed in a9In the Forex market, the standard lot size is 100, 000. A pip is the smallest price move that a given

exchange allows. Major currency pairs are priced to four decimal places; the smallest change is that

of the last decimal point. For example, for the currency EUR/USD, 1 pip is equivalent to $0.0001.

Therefore, for a trade of size $100,000, the platform charges followers $20 as commission.

12

follower account (that is $5 for a trade of $100,000). During our study period, the trader’s

compensation is calculated on a monthly calendar basis, but a trader is compensated only

if the profit earned in the trader’s account for that month is positive. Traders either get

fully compensated or zero for their trades in the month.

4 Empirical InvestigationsIn this section, we first describe how we measure the hypothetical profit (loss) of

executing a trade after some time delay is added to the original, real-time trade. We then

explore what factors influence the profit-gap, especially how magnitude of delay impacts

the profit-gap. Finally, we study what drives the amount of money following a trader

and empirically demonstrate the platform’s dilemma concerning delay.

4.1 Simulating Delayed Trades

Let us assume that a trader opens a particular trade at time t1 and closes this trade

at time t2; the difference (t2− t1), is called holding time. Such trade level data is directly

observed in this study. Also observed is the profit or loss associated with the trade. Now,

consider a hypothetical trade that is the same as the above trade (the same currency,

type, and standard lot size) but the open time and close time are, respectively, t1 + δ1,

and t2 + δ2. Here, δ1 and δ2 denote the open delay and close delay respectively. We

can recover the profit (loss) of this delayed trade using the historical spot prices in the

currency market traded at time t1 +δ1, and t2 +δ2. We vary the values for open and close

delay at different levels: 0, 5, 15, 30, 60, 90, and 120 minutes 10. When δ1 = δ2 = 0, it

represents a real-time trade. The open and close delay are chosen such that the simulated

close time is later than the simulated open time. Thus, we simulate a maximum of 48

distinct hypothetical trades corresponding to each original (or real-time) trade in the

data sample. Figure 2 depicts how we generate the hypothetical trade after considering

time delay.

For each simulated trade, we calculate the profit-gap as the real-time profit minus

the simulated profit. The profit-gap directly measures the economic value of acting upon

real-time information, relative to acting upon delayed information. The higher the profit-

gap value, the more worthwhile for followers to follow the real-time trade (with fees). We

would like to mention that the profit-gap value can be negative for some trades, meaning

the delayed trades may even earn higher profit than the real-time trades. However, we10A delayed trade of more than 2 hours is hardly worth considering in the volatile Forex market.

13

believe that a good trader is able to time the market (when to enter and exit the market).

Delaying the trade would then lower the value of her original trade.

Figure 2: Illustration of Generating Hypothetical Trade

4.2 What drives profit-gap

We next investigate how different factors affect the value of following the trade in

real-time. To this end, we specifically explore how the magnitude of delay influences the

profit-gap.

We use i to index a trade, j to index a trader, m to index a currency market, and t

to index each 15-minute time period across the 17 months 11. The dependent variable is

ProfitGapi,j,t. The explanatory variables of interest are OpenDelayi,j,t, CloseDelayi,j,t,

and Followersi,j,t. The variable OpenDelayi,j,t (CloseDelayi,j,t) is the open (close) delay

added for trade i, executed by trader j at time period t. The variable Followersi,j,t

represents the total number of followers following trader j at time t 12. The econometric

model we consider is specified as follows:

ProfitGapi,j,t = β1OpenDelayi,j,t + β2(OpenDelayi,j,t)2 + β3CloseDelayi,j,t+

β4(CloseDelayi,j,t)2 + β5Followersi,j,t +Xi,j,t + Traderj + Currencym + Timet + εi,j,t

(1)

In equation (1), Traderj is the trader-level fixed effects, Currencym is the currency

(market) specific fixed effects(13, and Timet is the time-specific fixed effects capturing the

market status corresponding to a specific time period. The term Xi,j,t represents a vector

of control variables, including the standard lot size of trade i (StandardLotsi,j,t), the

highest potential profit (measured in pips) during the holding time of trade i by trader11We use a relatively short time window of 15 minutes because of the high volatility of the Forex

market.12In the regression model, we take the logarithm of the followers to account for skewness in this

variable.13For notation simplification, we suppressed m from the subscription of all the other variables.

14

j at time t (HighestProfiti,j,t), the lowest potential profit (worst drawdown) during the

holding time of trade i by trader j at time t (WorstDrawdowni,j,t), the holding time

of trade i by trader j at time t (HoldingT imei,j,t), the total profit (with unit dollars)

of trader j by time t (Profiti,j,t), the total profit (with unit pips) of trader j by time

t (Profitpipsi,j,t), the highest profit of a trade among all the trades of trader j until

time t (BestTradei,j,t), the lowest profit of a trade among all the trades of trader j until

time t (WorstTradei,j,t), the number of trades of trader j by time t (Tradesi,j,t), the

annual return of investment of trader j by time t (ROIi,j,t), the percentage of trades

with positive profits among all the trades of trader j by time t (WinRatioi,j,t), the

number of weeks trader j has been traded on the platform by time t (Agei,j,t), the rank of

trader j at time t (Ranki,j,t), the average profit of trader j by time t (AvgTradei,j,t), the

maximum number of open trades trader j has been held by time t (MaxOpenTradei,j,t),

the minimum capital required to trade all the trades from trader j in a follower’s account

at time t (MinEquityi,j,t), the profit of open (unrealized) positions of trader j at time

t (OpenPositioni,j,t), the number of views trader j’s profile page has received by time

t (V iewi,j,t), and the indicator whether there are followers following trader j at time t

(DummyFolloweri,j,t). Table 2 tabbulates the summary statistics of profit-gap, open

delay, close delay, etc.

To test potential existence of multicollinearity, we calculate the variance inflation

factors (VIF) and find that they are well below the acceptable threshold (10), indicating

the absence of multicollinearity. We use the White test to check for heteroscedasticity,

and the result (the chi-square value is 1,467,927 with p-value less than 0.001) shows

significant heteroscedasticity in the error term. We therefore control for it using robust

standard errors.

We simulate hypothetical trades with different magnitudes of delay. Please note that

the volume of our hypothetical trades are minuscule relative to the gigantic Forex market.

It is unlikely that any trade in our data would be subject to instant market impact (slip-

page) 14. Therefore, we assume exogeneity of the variable OpenDelay and CloseDelay.

Potential endogeneity might arise in variable Followersi,j,t. For example, it can be ar-

gued that higher profit-gap may likely be the cause why followers follow the trader, i.e.14For example, the daily trading volume of Forex market is on average 1.8 trillion in January 2018.

The maximum amount of money following a trader in our data is 3 million dollars.

15

Table 2: Summary Statistics of DataVariable Obs Mean Std. Dev. Median Min Max

ProfitGap 48,125,024 0.69 19.79 0.30 -1,002.20 1,031.90

OpenDelay 48,125,024 31.30 37.15 15 0 120

CloseDelay 48,125,024 42.22 42.89 30 0 120

Followers 48,125,024 9.22 59.14 0 0 2,589

StandardLots 48,125,024 3.75 86.63 0.06 0.01 1e5

HighestProfit 48,125,024 18.79 24.35 11 0 881

WorstDrawdown 48,125,024 -19.66 22.56 -12 -1,351 0

HoldingT ime 48,125,024 190.74 250.64 95.35 1 1,440

Profit 48,125,024 144.95 21,221.78 1.35 -1.77e6 1.28e7

Profitpips 48,125,024 5.38 31.32 6.10 -655.60 772.40

AvgTrade 48,125,024 5.30 72.04 1.00 -10,611.30 4,898

MaxOpenTrade 48,125,024 45.94 49.99 30 0 428

Rank 48,125,024 9,826.72 12,034.6 2,598 1 32,001

Trades 48,125,024 1,888.94 2,353.35 911 0 14,676

Age 48,125,024 39.45 49.71 23 0 478

Amount 48,125,024 10,308.21 63,718.35 0 0 3,267,296

BestTrade 48,125,024 508.49 1,360.10 200 -499.60 69,938.40

WorstTrade 48,125,024 -785.72 8,291.01 -371 -1,807,046 0

WinRatio 48,125,024 65.45 22.57 68 0 100

ROI 48,125,024 392.28 3,261.81 32 -11,963 84,266.6

MinEquity 48,125,024 1,866.37 2,050.23 1,064 0 9,955.44

OpenPosition 48,125,024 -1,529.13 11,028.39 0.00 -1,486,899 622,768.30

V iew 48,125,024 8,912.46 30,339.39 1,834 0 2,258,260

DummyFollower 48,125,024 0.08 0.27 0 0 1

16

simultaneity may occur.

To account for potential endogeneity, we use the two-stage least squares (2SLS) re-

gression with instrument variables (IV). We construct two IVs for variable Followersi,j,t.

The first IV (IV1) uses the amount following other traders of the same followers. Specifi-

cally, IV1 is the Hausman type of instrument, constructed as the total number of followers

(from the set of followers following the focal trader) following other traders (but exclud-

ing the focal trader) at time t. Variable IV1 is correlated with Followersi,j,t because

it is constructed from the same set of followers; but it should not directly influence the

focal trader’s performance (the dependent variable ProfitGap) because of the exclusion

of the focal trader from the construction of IV1. Another IV (IV2) is the average rank

of other traders who are followed by the same set of followers at time t. The rank of

a trader is displayed by ZuluTrade. Likewise, the average rank of other traders should

be correlated with the number of followers because this is a metric followers care about

when deciding whether to follow a trader; but rank does not directly influence the focal

trader’s performance because it is constructed from the other traders (Rossi 2014).

To validate our IVs, we perform a Hausman test, where under the null hypothesis the

specified endogenous regressors can actually be treated as exogenous. We have chi-square

value 1190.6 with p-value < 0.001, indicating the preference of IV based estimation. The

IVs should further satisfy two prerequisites: the relevance assumption and the exogeneity

assumption (Green 2007). The relevance assumption requires that the IVs should be cor-

related with the endogenous variables and that this correlation should not be weak. The

F-statistics for the endogenous variable (followers) is 46,145 (p-value = 0.000). The other

condition, exogeneity, requires that instruments excluded from the structural equation

must be uncorrelated with the structural errors, which is typically done using a test of

over-identifying restrictions via Hansen’s J statistic (Hansen 1982). The Hansen J statis-

tic is 2.642 with p-value = 0.1041, indicating that we do not reject the null hypothesis

that the instruments are exogenous and excludable.

Table 3 presents the estimation results corresponding to the econometric model in

equation (1). Larger magnitudes of open and close delay increase the profit-gap, indicat-

ing that it is more valuable for investors to follow in real-time if the platform releases a

trader’s trade with longer delay. From Table 3, we see that the coefficients of the squared

delay terms are significantly negative. Thus, the profit-gap increases with delay in a con-

17

cave manner. Holding everything else equal, the marginal effect of delay on the profit-gap

decreases with larger delay magnitude. As the platform adds larger magnitude of delay,

the impact on profit-gap decreases. Besides, the profit-gap increases with larger number

of followers following a trader. The profit-gap is heterogeneous across different currency

markets, and it varies as the level of market volatility varies in different markets.

Table 3: The Impact of Delay on the Profit-Gap

Dependent Variable: ProfitGap

VARIABLES Coefficients

OpenDelay 0.016***(0.000)

(OpenDelay)2 -0.0009***(0.000)

CloseDelay 0.017***(0.000)

(CloseDelay)2 -0.0007***(0.000)

Followers 0.154***(0.001)

Control Variables Yes

Currency-level FE Yes

Trader-level FE Yes

Time-level FE Yes

Observations 48,125,024

R-squared 0.128

Note. The robust standard error is reported in parenthesis.

*** p < 0.01; ** p < 0.05; * p < 0.10.

4.3 Transparency-revenue Conundrum

To investigate what drives the amount following a trader, we estimate the regression

model as specified in equation (2). The dependent variable is the amount following trade

i from trader j at time t. The variables of interests are the number of (cumulative)

views of trader j received by time t (V iewi,j,t), the profit-gap (with 30-minute open

delay and zero close delay as implemented by ZuluTrade) of trade i from trader j at

time t (ProfitGapi,j,t), and the interaction term of view and profit-gap. The number of

views captures the potential number of interested users that the trader might be able to

convert into followers. From the results with respect to equation (1), we know that larger

magnitude of delay is associated with higher profit-gap. The profit-gap measures the

18

information transparency level. Xi,j,t represents the set of control variables as specified

in equation (1). We take logarithms of amount and view. We also control for the trader-

level and time-level fixed effects.

Amounti,j,t = θ1V iewi,j,t + θ2ProfitGapi,j,t + θ3V iewi,j,t ∗ ProfitGapi,j,t +Xi,j,t

+Traderj + Timet + εi,j,t

(2)

The variance inflation factors are all well below 10, indicating absence of multi-

collinearity. We control for heteroscedasticity using robust standard errors. Both vari-

ables V iew and ProfitGap might be subject to endogeneity. For example, it can be

argued that a trader’s profile attracts more views because the trader has higher amount

of money following her; followers choose to follow a trader because it is worthwhile to

follow her in real time (higher profit-gap), i.e. simultaneity may occur. Hence, we con-

struct instrument variables for both variables. The IV we construct for ProfitGapi,j,t is

the holding time of trade i. The profit-gap is sensitive to the holding time of the trade,

but a follower’s decision regarding whether to follow a trader or not does not directly rely

on the holding time of the trade.

We construct two IVs for variable V iewi,j,t. The first IV (IV1) uses sum of the number

of views of the other traders followed by the same set of followers following the focal trader

j. Specifically, IV1 is the Hausman type of instrument, constructed as the total number

of views (from the set of followers following the focal trader) following other traders

(but excluding the focal trader) at time t. Variable IV1 is expected to correlate with

V iewi,j,t because it is constructed from the same set of followers; but it should not be

systematically correlated with the focal trader’s performance (the dependent variable

Amounti,j,t) because the focal trader is excluded from the construction of IV1. The other

IV (IV2) is the average rank of other traders who are followed by the same set of followers

at time t. The rank of a trader is displayed by ZuluTrade; traders with higher ranks are

listed at more prominent positions, attracting more views. Likewise, the average rank of

other traders should be correlated with the views because this is a metric followers care

about when deciding whether to follow a trader; but rank does not directly influence the

focal trader’s views because it is constructed on the other traders (Rossi 2014).

The Hausman test statistics is 864.6 with p-value < 0.001, indicating the preference of

IV based estimation. The F-statistics for the three variables (ProfitGap, V iew, and the

19

interaction term) are 8,144.60 (p-value < 0.01), 2,246.30 (p-value < 0.01), and 17,621.87

(p-value < 0.01) respectively, pointing to strong IVs. The Hansen J statistic is χ2(1) =

1.647 with p-value = 0.199, satisfying exogeneity condition.

Table 4: The Driving Forces of the Amount Following a Trader

Dependent Variable: Amount

VARIABLES Coefficients

View 3.281*** ( 0.055 )

ProfitGap 0.289*** ( 0.095 )

View*ProfitGap -0.031*** ( 0.01 )

Rank -0.003*** ( 0.000 )

AgeLog -4.055*** ( 0.081 )

BestTrade 0.001** ( 0.000 )

WorstTrade -0.001*** ( 0.000 )

WinRatio 0.052*** ( 0.002 )

ROI 0.001*** ( 0.000 )

Trades -0.002*** ( 0.000 )

AvgTrade 0.001*** ( 0.000 )

MinEquity -0.001*** ( 0.000 )

StandardLots 0.002 ( 0.000 )

HighestProfit -0.005*** ( 0.001 )

WorstDrawdown -0.005*** ( 0.002 )

Profit 0.001 ( 0.001 )

MaxOpenTrade -0.008*** ( 0.001 )

Profitpips 0.003*** ( 0.001 )

OpenPosition 0.001*** ( 0.000 )

Trader-level FE Yes

Time-level FE Yes

Observations 730,812

R-squared 0.912

Note. The standard error is reported in parenthesis.

*** p < 0.01; ** p < 0.05; * p < 0.10.

20

The results in Table 4 show that both coefficients of view and profit-gap are significant

and positive. Interestingly, the interaction term of the two is negative, implying that view

and profit-gap weaken each other. The marginal effect of profit-gap on the amount is

(0.289 − 0.031 × V iew). When the number of views of a trader attracted is low, higher

profit-gap is associated with higher amount following the trader. However, when the

number of views is large, higher profit-gap results in lower amount following the trader.

In other words, in the case when a large number of potential followers are interested

in a trader, low transparency would make it more difficult for potential followers to

evaluate the trader’s ability and thus reduces the chance of converting them into followers.

The marginal effect of view on the amount is (3.281 − 0.031 × ProfitGap) and can be

interpreted in a similar manner. Taken together, this negative interaction presents the

empirical evidence that the platform faces the trader-off between increasing transparency

(to facilitate performance evaluation for interested investors and help traders convert

those potential followers) and reducing transparency (to prevent free-riding). Having

demonstrated the tradeoff, we next devise various information release policies.

5 Optimizing Information ReleaseIn this section, we develop and compare several information release policies that handle

the transparency-revenue tradeoff in different ways. Central to the manner in which the

tradeoff is handled, is the notion of Money-at-Risk, a concept we operationalize later

in this section. This concept was introduced to us during our interactions with the

ZuluTrade management team. It is important to mention that the policies we study

in this section (including the current release policy employed by ZuluTrade), do not

explicitly optimize the revenue of the platform. Rather, these policies can be expected to

indirectly impact the revenue in a desirable manner. Direct revenue optimization policies

are introduced in the next section.

Conceptually, the platform would like to maximize the transparency of each trade in

order for potential followers to evaluate traders, without endangering its revenue collected

from commissions and follower fees. Thus, the platform needs to strike a fine balance

between transparency and its revenue objectives. Transparency is maximized when a

trader’s trades are released to the public without any delay. However, such a (complete)

transparency policy runs the risk that followers might manually copy the trades of a

trader and execute them elsewhere without paying follower fees. This suggests the notion

21

of Money-at-Risk, defined as the expected loss in the amount following a trader for a

given delay and follower fee. To put it simply, Money-at-Risk (MaR) measures the risk

of losing the amount following a trader. Mathematically, the Money-at-Risk for a trader

is the amount following the trader (ai) multiplied by the probability that the profit-gap

of the trades associated with this trader is less than the follower fee (c) or,

aiPi{xi(δ) ≤ c},

where xi(δ) is the random variable representing the profit-gap of the trades associated

with a trader. From the platform’s perspective, an important goal is to limit the total

Money-at-Risk across all the traders, namely,∑A

i=1 aiPi{xi(δ) ≤ c}, where A 15 is the

total number of traders.

We first introduce a theoretical model to capture how the profit-gap of a trade evolves

as a function of the delay. This model will be proven useful in the analysis of the

information release policies that are studied in this section.

5.1 Profit-Gap Model

Based on the empirical findings in Table 1, the mean profit-gap was observed to be

a concave increasing function of the delay. This finding suggests the following stochastic

process to model the profit-gap of a trade (xi(t)) after introducing a delay of t from the

point in time when the real-time trade was executed by trader i.

dxi(t) =αi

2√tdt+ σidz(t),

where the parameters αi and σi are trader specific.

The above model (VABM) is a variant of an Arithmetic Brownian Motion (ABM),

where the drift term is inversely proportional to the square root of time (t). This model

can be solved to show that the value of xi(t) is Normally distributed with a mean of αi√t

and a variance of σ2i t. The mean αi

√t, reflects the concave, increasing relationship with

delay.

We apply Maximum Likelihood Estimation (MLE) to estimate the parameters for the

profit-gap of a trader (α̂i and σ̂i for trader i). The VABM model fits the data better15More precisely, we should interpret A as the total number of trader accounts. For convenience,

however, we will use the term trader to mean trader account. These two terms differ in practice, but

inconsequential to our mdoel.

22

than two alternative models: (1) a model with no drift term, or Brownian Motion (BM),

and (2) a model with a linear drift term, or Arithmetic Brownian Motion (ABM). The

average Bayesian Information Criterion (BIC) values for a sample of of 1,000 traders are:

VABM (85,200), ABM (86,600) and BM (87,600), showing that VABM has the best fit

among the three models. This finding shows that, generally speaking, the profit-gap of

a trade evolves with delay with a positive and concave drift, not just as white noise or

with linear drift. However, there are indeed some traders with negative drift and some

others with zero drift.

For any given trader and delay δ , we can use the above VABM process to find the

probability that the profit-gap associated with the trades of the trader is less than the

following fee c. This probability is calculated as the cumulative distribution function of

the normal distribution with mean α̂i√δi and variance σ̂2

i δi, Fi(c|δi), where c = 1, α̂i

and σ̂i are the VABM parameters for trader i. Figure 3 illustrates how the probability

changes with delay for a representative trader with id 102885. We see that the fitted

probability F̂i(c|δi) decreases with δi in a convex manner. This convex property is seen

to hold for most traders.

Figure 3: The fitted probability with delay for trader 102885

5.2 Comparison of Different Release Policies

We first propose an indifference information release policy that maximizes trade trans-

parency while ensuring that the release of trade information is such that the expected

profit-gap for a trader is exactly equal to the following fee. This policy is based on the

23

premise that, on average, there will be no loss of commission revenue because the release

of trade information is such that a risk neutral follower will be indifferent between fol-

lowing in real time (with a fee) and executing the same trade with delay (for free). Next,

we propose two policies that consider the profit-gap distribution to decide on the delay,

rather than basing this decision on the mean of this distribution. The first is a uniform

release policy where all trades are publicly posted for everyone to see with a delay of

δ. The second policy is a customized money-at-risk release policy where the information

release policy (i.e., delay, δi) depends on the characteristics of trader i. The uniform

policy is, of course, a subset of the customized money-at-risk release policy.

5.2.1 Customized Indifference Policy

In this policy, a trader’s trades are released with a delay such that the expected profit-gap

equals the difference in commission fees between real-time and delayed trades.

Let Fi(x|δi) (fi(x|δi)) denote the cumulative probability density (probability density

function) for the profit-gap (x) associated with trader i when the trades of this trader are

released with a delay of δi. Because the profit-gap could be negative, we allow the support

of the profit-gap distribution to be x ∈ [-∞, +∞]. Then, the customized indifference delay

would be given by the value of δi such that∫ +∞

−∞xfi(x)dx = c,

where c is the fee paid to follow the trades of a trader.

5.2.2 Customized Money-at-Risk Policy

In this policy, each trader (or specifically, each trader’s account) can be customized with

a different release delay, δi. The optimization problem can be stated as below.

Minimizeδi,i∈{1,2,...,A}

A∑i=1

viδi

subject toA∑i=1

aiFi(c|δi) ≤ η

In the above specification, δi is the decision variable, A is the total number of traders,

ai is the amount of money following trader i, vi is the number of views received by trader

i in a period, Fi(c|δi) is the probability the profit-gap less than the following fee c given

24

delay δi, and η is maximum money-at-risk (the platform can tolerate). Intuitively, the

platform should release a trader’s trades with less delay if the trader receives more views,

to help the trader convert more potential followers.

In the above problem, the objective function is a measure of trade opacity ; thus we

wish to minimize this measure. Our data tracks the number of potential followers view-

ing a trader. We denote the total number of page views per period as (vi). These page

views are a direct measure of potential followers evaluating the trader. The constraint in

the above problem represents the Money-at-Risk (MaR). When the probability that the

profit-gap for a trader is below the follower fee (Fi(c|δi)), investors who follow the trader

could switch from following in real-time to free-riding with some delay (δi). This prob-

ability is a measure of the vulnerability of losing the commissions the platform receives

from the followers of a particular trader. The Money-at-Risk captures this vulnerability,

and is calculated as the probability Fi(c|δi) multiplied by the amount following.

The probability Fi(c|δi) decreases with δi. At the extreme, if δi = 0, the real-time

profit and the delayed profit are the same; hence, the profit-gap is zero. Thus, the

probability that the profit-gap is less than c is 1. As the delay increases, the delayed

profit reduces (while the real time profit stays the same); hence the profit-gap increases.

Thus, F ′i (c|δi) < 0. Based on our regression results (see Table 3), we would expect the

profit-gap to exhibit diminishing returns with respect to delay, implying that the profit-

gap can be expected to be a concave function of the delay. Thus, we can expect that the

cumulative density F to be convex with respect to the delay, i.e., F ′′i (c|δi) > 0, a property

that is also supported in data.

The function Fi(c|δi) is trader specific and embeds in its knowledge of how proficient

the trader is at spotting short-lived opportunities in the market. For the same delay

(say δ1), if Fi(c|δ1) < Fj(c|δ1), it implies that trader i′s trades carry more information

value, or equivalently, for the same delay, these trades are more worthwhile to follow in

real time. Thus, they can be released earlier to serve the goal of increasing transparency.

However, relative to trader i, trader j′s trades need more protection (from the perspective

of information value), and need to be released later.

To solve the above problem, it is sufficient to note that it is a convex optimization

problem. This is because the objective function is linear in the decision variables while

the constraint is jointly convex in the decision variables. In the constraint, it is easy to

25

see that the decision variables do not interact with one another and the data supports

the property that the cumulative density associated with trader i, Fi(c|δi), is convex in

δi.

5.2.3 Uniform Policy

Minimizeδ

A∑i=1

viδ

subject toA∑i=1

aiFi(c|δ) ≤ η

As before, the objective function is a measure of trade opacity ; thus we wish to

minimize this measure. The constraint is also conceptually similar. The uniform policy

is a special case of the customized money-at-risk policy where all trader’s trades are

released with the same time delay (δ).

5.2.4 Numerical Illustration

We randomly choose ten traders to illustrate how the different release policies work. The

amount of money following each trader (ai) is [2800, 400, 126, 773, 150, 58, 324, 650, 58, 200].

The average views per period (day) for each trader (vi) is [1, 1, 1, 2, 3, 3, 3, 3, 5, 8]. We take

the following fee as c = 1 (with unit pips). For each trader, we first estimate the param-

eters αi and σi using the data and obtain the cumulative probability function Fi(c|δi).

Given different values of the maximum money-at-risk (η), we can calculate the optimal

opacity and draw the Pareto curve associated with the uniform policy as shown in Figure 4

(a). The uniform 30-minute delay policy (marked as point U1) adopted by the ZuluTrade

platform corresponds to a money-at-risk of 1694.8 and an opacity measure of 900. Given

the same money-at-risk value, the optimal opacity is only 460.9 under the customized

money-at-risk release policy (marked as point C1). The optimal delay for each trader

is [91.2, 31.7, 24.5, 32.9, 3.3, 4.6, 28.7, 19.1, 3.3, 8]. As can be seen from the data, the first

three traders have the same number of views (v1 = v2 = v3 = 1), but the amounts

of money following these traders are different. It is better to use a longer delay for a

trader with a higher amount of money following. This is done to protect the revenue

earned from the current followers of such traders. We can see that a1 > a2 > a3 results

in δ∗1 > δ∗2 > δ∗3. Conversely, everything else held constant, for a trader with a higher

26

number of views, it is better to use a shorter delay.

Figure 4: The Comparison of Different Release Policies

In Figure 4 (a), point I1 represents the customized indifference policy with a money-

at-risk value of 2769.5 and an opacity of 380. The delay added for each trader’s trades

is [0.4, 26.6, 1.9, 47.4, 2.1, 4.4, 11.3, 0.4, 1.6, 24.2]. Given the same level of money-at-risk,

the optimal opacity is 28.8 under the uniform policy with 1 minute delay (point U2), and

16.3 under the customized money-at-risk policy (point C2). The optimal delay for each

trader is [4.5, 1.8, 0.5, 1.3, 0.2, 0, 1.1, 1, 0, 0].

In Figure 4 (a), we can see that the uniform release policy performs better than

the customized indifference policy in terms of opacity given the same level of maximum

money-at-risk. However, these two policies do not have a strict ordering in terms of their

opacity values for the same money-at-risk. For example, as illustrated in Figure 4 (b), for

different set of traders, the customized indifference policy (point I2) can perform better

than the uniform policy (point U3). Point C3 represents the customized money-at-risk

policy.

It is not surprising that the customized money-at-risk release policy always dominates

the uniform release policy, since the uniform policy is a subset of customized money-at-

risk policy. However, the customized money-at-risk policy can significantly outperform

the uniform policy given the same money-at-risk.

6 Revenue OptimizationSo far, we have considered information release policies that attempted to strike a

balance between information transparency (represented as views-weighted delay) and

revenue (represented as Money-at-Risk). Our optimization formulations for these poli-

27

cies, however, do not directly consider platform revenue as the objective to optimize.

We next present a stochastic control model to directly optimize platform revenue. This

model is developed at the trader level, i.e., for every trader, depending on the trader’s

parameters, we implement the feedback control δ(a(t), v(t)), representing the delay intro-

duced for a particular trader when the current amount following is a(t) and the number

of accumulated views is v(t). Because we pose an infinite horizon model, we do not need

to explicitly consider time (t) in the choice of the control, although the variables a and v

are, of course, functions of time.

The objective of the platform would be to maximize its revenue – from commissions

and following fees – that is generated from the amount following a trader, a(t). The

revenue per unit time (or the revenue rate) earned by the platform can be expected to

increase with the amount. Thus, we model the revenue rate as k times a(t), where k is a

trader specific constant that could be chosen based upon the frequency of trades executed

by the trader. As mentioned above, there are two state variables that drive the feedback

control, a(t) and v(t). We accordingly propose two stochastic differential models for the

evolution over time of both state variables.

The amount following the trader is modeled using a stochastic differential equa-

tion whose drift term is a function of the control δ(a(t), v(t)) and the views v(t), as

g(δ(a(t), v(t)), v(t)). This is consistent with our understanding of the forces that drive

the amount following a trader. For any given trader, we expect that the number of views

generates an inflow into the set of current followers. However, this inflow is throttled if

the opacity increases, i.e., the recent investment activities of the trader are not visible to

potential followers. On the other hand, increasing transparency runs the risk of losing

current followers. The noise term is modeled as σadZ(t), where σa is the volatility coef-

ficient and dZ(t) is the increment of the Wiener process following N (0, dt). Following

convention in control theory, we suppress the time argument and write the state equation

for the amount as below.

da(t) = g(δ, v)dt+ σadZ(t)

Since ZuluTade uses a one-size-fits-all uniform release policy (δ = 30 minutes), the

impact of the control on the change in the amount following a trader cannot be empirically

observed. However, concerning the impact of delay, we expect that the inflow of new

followers to be a function of the delay whereas the outflow of current followers to depend

28

on the square root of delay. The influence of the delay on the inflow is based on the

assumption that new followers cannot evaluate a trader’s current trades if the delay is

large. On the other hand, the outflow should depend on the square root of delay, rather

than directly on the delay. This is because the square root of delay is proportional to the

profit-gap (as we show in Section 5.1), a quantity that ultimately controls a follower’s

decision to pay the follower fee and execute the real-time trade, or execute the delayed

trade released by the platform for free.

We therefore propose that the drift term (or the mean change in the amount following

a trader) to be given by

g(δ, v) =

(−αv

√a

c+ δ+ γ√v

a

c+ δ

)We expect α, γ > 0. The first term in the expression for g(δ, v) represents the outflow

in the amount; increasing δ reduces the outflow since it protects the current amount of

money following the trader. We take the square root of delay because profit gap, rather

than delay, can be expected to influence the outflow. The outflow should increase with the

amount, but do so in a diminishing manner: As the amount following a trader increases,

it could slow down the increase in the outflow. Finally, the outflow can be expected to

increase with the number of views. This can be understood as a learning effect. The

exit decision of a follower depends on a comparison between the profit gap (a stochastic

quantity) and the following fee (a deterministic quantity). Thus before leaving a trader,

a follower would likely need to learn how the profit gap for the trader compares with the

following fee. To capture this learning effect, we use the number of views v to represent

the amount of learning that has occurred for the trader.

The second term in the drift is the inflow in the amount following a trader. This term

captures the effect that a higher delay makes it harder (for followers) to evaluate a trader;

this effect is aggravated if the number of views and the amount is also high. The amount

following a trader should influence the level of interest for the trader: a higher amount

means that the trader can be expected to attract more attention from potential followers.

Similarly, the number of views should also generate more interests in the trader’s profile

page. However, the amount effect can be expected to be stronger – hence, the square

root operator is applied to the views but the amount effect is linear.

The views received by the trader is modeled as another stochastic differential equation

29

whose drift term is a function of the amount following and the current number of views,

h(a, v). The noise term is modeled as σvdW (t), where σv is the volatility coefficient and

dW (t) is the increment of the Wiener process. Thus we have,

dv(t) = h(a, v)dt+ σvdW (t)

Conspicuously absent in the above relationship is the impact of the control variable

(delay). However, this is expected. While the ability of potential followers to evaluate a

trader depends on the delay, we do not expect the control to affect the evaluation interest

(or views) from potential followers. Because we do not expect the control (delay) to affect

the change in views in a direct manner, we look for empirical support for the function

h(a, v). Turning to the data, it is natural to expect that a higher amount following

(number of views) generates more interest. In Table 5, we find empirical support for a

relatively simple form for the change in the number of views: ∆v(t) = v(t+ 1)− v(t) =

c3a(t) + c4v(t), by running the regression in equation (3). Therefore, empirical evidence

allows us to propose h(a, v) = pa+ qv.

∆V iewj,t = θ1Amountj,t + θ2V iewj,t +Xj,t + Traderj + Timet + εj,t (3)

Table 5: Impact of Amount and Views on the Change in Views

Dependent Variable: ∆V iew

VARIABLES Coefficients

Amount 0.003***(0.0003)

V iew 0.005***(0.0006)

Control Variables Yes

Trader-level FE Yes

Time-level FE Yes

Observations 181,666

R-squared 0.020

Note. The standard error is reported in parenthesis.

*** p < 0.01; ** p < 0.05; * p < 0.10.

6.1 Stochastic Control Model

Based on the above discussion, we present the revenue optimization problem for the

platform as the following stochastic control model, after suppressing all implicit time

30

arguments.

maxδ

E[∫ T

0

kae−ρtdt

]subject to da =

(−αv

√a

δ0 + δ+ γ√v

a

δ0 + δ

)dt+ σadZ

dv = (pa+ qv)dt+ σvdW

(4)

where δ is the control parameter, a is the amount of money following a trader at time t,

v is the number of views at time t, and ρ is the discount factor.

6.1.1 Solution

To solve the stochastic optimal control problem, the optimal control δ should satisfy the

following Hamilton-Jacobi-Bellman (HJB) equation for the value function V .

ρV (a, v, t) = maxδ{ka+ Va

(−αv

√a

δ0 + δ+ γ√v

a

δ0 + δ

)+ Vv(pa+ qv) + Vt

+1

2Vaaσ

2a +

1

2Vvvσ

2v}

(5)

Using the above equation, a point-wise optimization problem can be solved as below.

The optimal solution is bang-bang, implying that the trades of a trader should either

be released with minimum possible delay (δ0) or the highest delay that is reasonable. In

this context, we can set δ0 to be a small value (say a few seconds) to represent the min-

imum possible delay that can be implemented given information technology constraints.

Concerning the maximum delay, we can set this value to a relatively large value, e.g.,

120 minutes. A transaction released beyond 120 minutes would likely be independent of

the original real-time transaction and increasing the delay further would not make sense.

The above values of minimum possible delay and maximum delay are illustrative, and

could change with the platform’s technology and the market. Note that, the optimal

delay changes dynamically in the life of a trader, depending on the current values of the

state variables associated with the trader (amount and number of views).

To prove the optimal policy structure, let g(δ) denote the terms involving the control

variable in the HJB as shown below.

g(δ) = −αv√

a

δ0 + δ+ γ√v

a

δ0 + δ

The first order condition yields

31

g′(δ) =

√av

2(δ0 + δ)32

(α√v − 2γ

√a√

δ0 + δ

)The optimal solution is a corner solution as shown below.

• If g′(0) ≥ 0, then g(δ) increases in δ. Hence, the maximum delay should be chosen.

Thus, δ? = δm, where (δ0 + δm) represents the maximum delay.

• When g′(0) < 0, then the optimal solution is either δ? = 0 or δ? = δm. This can be

seen by setting g′(δ) = 0. Solving for δ = δ1 yields a minimum because g′′(δ1) > 0.

The value of δ1 is unique because once g′(δ) > 0, it stays positive. Hence, the

optimal solution can be found by evaluating the value of the function g(δ) at the

two extreme points (0,δm) and choosing the higher of the two values.

While the above bang-bang policy does not easily lend itself to an analytical solution

for the value function, a numerical value of the optimal discounted revenue for a trader

can be obtained for given values of a and v and the other trader specific parameters.

Currently, ZuluTrade provides a ranking of traders, whose ranking algorithm is not made

public. However, this ranking can be assumed to be based on trader characteristics such

as profit, risk, and so on, that signal the quality of the trader to potential followers.

It could also be based on factors that are in the interest of the platform, such as the

amount following. The numerical value of the optimal revenue can be incorporated into

the current ranking algorithm to provide a systematic way to infuse the economic goals

of the platform into the ranking of traders.

To illustrate the impact of the optimal policy on the platform’s goal, we compare,

for a hypothetical trader, the current policy (30 minute delay) and the bang-bang policy.

Figure 5 show how the amount evolves over time under the 30-minute policy and the

bang-bang policy. We see that the bang-bang policy outperforms the 30-minute policy in

terms of the amount of money. We also compare the value function (objective function

in equation (4)) by numerically calculating the integral. The value function is 1.11 under

the 30-minute policy and 1.18 under the bang-bang policy.

7 ConclusionWe studied information release strategies in social trading using data from ZuluTrade

that primarily engages in the Foreign Exchange (currency trading) market. The data

32

Figure 5: The Comparison between 30-minute and the bang-bang policy

Notes: α = 0.000001, γ = 0.000001, p = 0.000015, q = 0.00001, σa = 0.01, σv = 0.01,

ρ = 0.000000093, k = 2.6e− 7, T = 43200, dt = 1, a(0) = 100, and v(0) = 20.

tracks 15,352 traders, copied by 15,492 followers across five major currencies. Our main

interest was to determine the optimal information release policy for the platform. To-

wards this goal, we estimate the hypothetical profit earned by a trade if it was executed

with some delay (i.e., a virtual, simulated trade constructed by adding some time delays

to the open and close time of the original trade). To estimate the profit earned by a

simulated trade, we used historical spot price data of the currency being traded. Having

estimated the profit earned from the simulated trade, we developed a measure called

profit-gap, defined as the real-time profit minus the simulated profit. We estimated var-

ious econometric models to unravel the drivers of the profit-gap, the amount of money

following a trader, and the number of views received by the trader. Using the empirical

evidence gathered from the ZuluTrade platform, the study devised different ways for the

platform to address the transparency-revenue dilemma, both implicitly (by maximizing

transparency while respecting a Money-at-Risk constraint) and explicitly (by directly

optimizing the revenue generated from commissions and follower fees).

The key contribution of our proposed approach was in its ability to address the prob-

33

lem at the level of the trader, or at an even more granular level, such as a trader-currency

combination. This is in sharp contrast to the current approach of the platform that

adopts a simple, one-size-fits-all policy that chooses the same level of delay for all trades

occurring on the platform. Like the platform, our focus was on the use of delay as a

control variable, while holding the follower fee at a fixed level. A natural extension would

be to consider both the delay and the follower fee as control variables that are trader

specific. The inclusion of customized follower fees would provide a finer degree of control

for the platform: Rather than running the risk of losing existing followers, the platform

could offer a lower follower fee to maintain the existing set of followers associated with

a trader. The platform could also consider offering a menu of options for followers, each

with a given level of delay and an associated follower fee. The current options offered by

the platform lie at the two extremes: zero delay and a follower fee versus a 30-minute

delay and zero follower fee.

While the focus of our study was on the goals of the platform, our study can also

help followers answer the question whether they are paying too much for financial advice.

To illustrate, consider a follower that faces an uncertain profit-gap (x) that is Normally

distributed as N(α√δ, σ2δ). A natural question is: is paying a follower fee of c (per trade)

justified? As a simple answer, a comparison could be made between the expected profit-

gap (α√δ) and the following fee c: If α

√δ > c, then real-time following is worthwhile.

At a broader level, our study is one of the first of its kind to explore the value of

information in a real-world setting. The context in this study, namely the Forex market,

provided us with an ideal testbed to study information value measured by the profit

impact of delaying a trader’s trades. This is because the Forex market is very speculative

and volatile, and the impact of a small delay on the profit of a trade can be significant. In

other markets, e.g., the traditional stock market where the fundamentals do not change

rapidly, the impact of a small delay on profit may be too small to measure. Thus, to

study the impact of delay in the stock market, it will be necessary to simulate much larger

delays. The problem with introducing large delays, of course, is that exogenous events

that occur during the delay period could affect the profit-gap. Even so, the problem of

following in real-time versus doing so with delay is a prevalent phenomenon, not unique

to the Forex market.

In this study, we discussed the question of how to follow (i.e., real-time versus delay)

34

and the platform should manage delay in a way to maximize its revenue from commissions

and following fees. Future work could explore the additional question of whom to follow.

That is, it is possible that a trader-currency may not be worthwhile to follow at all,

whether in real-time or with delay. Together, the joint decision of whom to follow, and if

so, how to follow, is what must eventually be addressed for a more comprehensive study

of the problem.

35

ReferencesManuel Ammann and Nic Schaub. Social interaction and investing: Evidence from an

online social trading network. Working Paper, 2016.

Thomas Brand. What’s going on in the social trading and investing space, 2017.

URL http://blog.thomasbrand.xyz/2017/10/19/what-does-social-trading-

and-investing-mean-in-practice/. [Online; accessed 3-Jun-2018].

Bastian Breitmayer, Mona Mensmann, and Matthias Pelster. Social recognition and

investor overconfidence. Working Paper, 2017.

BrokerNotes. The modern trader: A report on the 13.9 million online traders, 2018. URL

https://brokernotes.co/modern-trader/. [Online; accessed 9-Oct-2018].

BusinessWire. Zulutrade continues to innovate adding social media features to its

website, 2017. URL https://www.businesswire.com/news/home/20170125005795/

en/ZuluTrade-Continues-Innovate-Adding-Social-Media-Features. [Online; ac-

cessed 1-Aug-2018].

Philipp Doering, Sascha Neumann, and Stephan Paul. A primer on social trading

networks–institutional aspects and empirical evidence. 2015.

Richard Eldridge. How social media is shaping financial services, 2017. URL

https://www.huffingtonpost.com/richard-eldridge/how-social-media-is-

shapi_b_9043918.html. [Online; accessed 9-Oct-2018].

FX Empire. The forex industry in 2018: Regulation, market size, and the future,

2017. URL https://www.fxempire.com/education/article/the-forex-industry-

in-2018-regulation-market-size-and-the-future-447777. [Online; accessed 3-

Jun-2018].

eToro, 2017. URL https://masterinvestor.co.uk/wp-content/uploads/2017/04/

MI_2017_Yoni_Assia_-_eToro.pdf. [Online; accessed 9-Oct-2018].

Florian Glaser and Marten Risius. Effects of transparency: analyzing social biases on

trader performance in social trading. Journal of Information Technology, 33(1):19–30,

2018.

36

William H Green. Econometric analysis (7 th), 2007.

Lars Peter Hansen. Large sample properties of generalized method of moments estimators.

Econometrica: Journal of the Econometric Society, pages 1029–1054, 1982.

Seung Hee Choi and Maneesh Chhabria. Effective delays in portfolio disclosure. Journal

of Financial Regulation and Compliance, 20(2):196–211, 2012.

Rawley Z Heimer. Peer pressure: Social interaction and the disposition effect. The Review

of Financial Studies, 29(11):3177–3209, 2016.

Michal Herzenstein, Scott Sonenshein, and Utpal M Dholakia. Tell me a good story and i

may lend you money: The role of narratives in peer-to-peer lending decisions. Journal

of Marketing Research, 48(SPL):S138–S149, 2011.

Reshma Kapadia. The fund industry is overdue for change, 2018. URL https:

//www.barrons.com/articles/fund-industry-1538767780. [Online; accessed 9-Oct-

2018].

P. Kim. Average mutual fund return, 2017. URL https://www.creditdonkey.com/

average-mutual-fund-return.html. [Online; accessed 9-Oct-2018].

Vanessa Kortekaas. Social trading targets savvy retail investors, 2013. URL https:

//www.ft.com/content/7fbe1624-d9bf-11e2-98fa-00144feab7de. [Online; accessed

1-Jun-2018].

In Lee and Yong Jae Shin. Fintech: Ecosystem, business models, investment decisions,

and challenges. Business Horizons, 61(1):35–46, 2018.

Woonyeol Lee and Qiang Ma. Whom to follow on social trading services? a system

to support discovering expert traders. In Digital Information Management (ICDIM),

2015 Tenth International Conference on, pages 188–193. IEEE, 2015.

Ulrike Malmendier and Devin Shanthikumar. Do security analysts speak in two tongues?

The Review of Financial Studies, 27(5):1287–1322, 2014.

Oliver Mäschle. Which information should entrepreneurs on german crowdinvesting-

platforms disclose? Technical report, Thünen-Series of Applied Economic Theory,

2012.

37

Andreas Oehler, Matthias Horn, and Stefan Wendt. Benefits from social trading? empir-

ical evidence for certificates on wikifolios. International Review of Financial Analysis,

46:202–210, 2016.

Wei Pan, Yaniv Altshuler, and Alex Pentland. Decoding social influence and the wisdom

of the crowd in financial trading network. Privacy, Security, Risk and Trust (PAS-

SAT), 2012 International Conference on and 2012 International Confernece on Social

Computing (SocialCom), pages 203–209, 2012.

Matthias Pelster and Annette Hofmann. About the fear of reputational loss: Social

trading and the disposition effect. Working Paper, 2017.

Reuters. Israeli social trading firm etoro raises $100 million in private funding,

2018. URL https://www.reuters.com/article/us-tech-etoro-fundraising/

israeli-social-trading-firm-etoro-raises-100-million-in-private-

funding-idUSKBN1GZ15S. [Online; accessed 1-Aug-2018].

Florian Röder and Andreas Walter. What drives investment flows into social trading

portfolios? Working Paper, 2017.

Peter E Rossi. Even the rich can make themselves poor: A critical examination of iv

methods in marketing applications. Marketing Science, 33(5):655–672, 2014.

Marno Verbeek and Yu Wang. Better than the original? the relative success of copycat

funds. Journal of Banking & Finance, 37(9):3454–3471, 2013.

Zhijin Zhou, Shengsheng Xiao, Yi-Chun Chad Ho, and Yong Tan. The persuasive and

informative effects of information disclosure: Evidence from an online supply chain

finance market. 2018.

38

Appendix

Table 6: VIF Values of Equation (2)

Variable VIF 1/VIF

V iew 4.18 0.24

Profitpips 4.18 0.24

AgeLog 3.25 0.31

HighestProfit 3.13 0.32

AvgTrade 2.14 0.47

WorstDrawdown 2.08 0.48

Trades 2.08 0.48

WorstTrade 1.96 0.51

MaxOpenTrade 1.94 0.52

MinEquity 1.67 0.60

BestTrade 1.51 0.66

WinRatio 1.37 0.73

ProfitGap 1.16 0.86

ROI 1.13 0.89

OpenPosition 1.11 0.90

Rank 1.09 0.92

StandardLots 1.01 0.99

Profit 1.01 0.99

39